🐛 [Bug] Torch-TensorRT fails at runtime when a symbolic dimension is derived from an int64 tensor and later reused in reshape after an aten.index.Tensor gather-style op.

[wenbingl]

Bug Description

Torch-TensorRT fails at runtime when a symbolic dimension is derived from an int64 tensor and later reused in reshape after an aten.index.Tensor gather-style op.
A minimized repro crashes with:

• RuntimeError: [Error thrown at core/runtime/execute_engine.cpp:249] Expected nbNames == 0 to be true but got false
• TensorRT shape error:
  IShuffleLayer ... reshaping failed ... Reshape dimension of -1 has no solution ... reshape dims{379350532 1 -1 2}

This matches the failure pattern seen in the larger ASR decoder FX graph (index -> _reshape_copy path with integer index tensors). And if the graph was compiled with a use_python_runtime flag, the test case runs ok.

To Reproduce

Steps to reproduce the behavior:

1. Save this repro as repro_issue.py:

import torch
import torch_tensorrt
import traceback

class TestModule(torch.nn.Module):
    def forward(self, x, targets, cache_length):
        # symbolic size from int64 tensor
        B = targets.size(0)
        # int64 tensor used as index
        idx = cache_length + torch.arange(1, device=x.device)
        y = x[:, idx, :]
        # reshape uses symbolic dim from int64-derived size
        z = y.reshape(B, 1, -1, 2)
        return z

B, S, D = 16, 128, 1024
x = torch.randn(B, S, D).cuda()
targets = torch.randint(0, 10, (B, 1), dtype=torch.int64).cuda()
cache_length = torch.tensor(0, dtype=torch.int64).cuda()

torch._dynamo.mark_dynamic(targets, 0)
torch._dynamo.mark_dynamic(x, 0)

model = TestModule().eval().cuda()

try:
    compiled_model = torch.compile(
        model,
        backend="tensorrt",
        options={
            "truncate_double": True,
            "enabled_precisions": {torch.float, torch.half},
            "min_block_size": 1,
            "optimization_level": 1,
            "debug": True,
        },
    )
    compiled_model(x, targets, cache_length)
except Exception:
    traceback.print_exc()

2. Run:
   python repro_issue.py

3. Observe runtime failure in TensorRT execution with inferShapes/reshape error and Expected nbNames == 0 ... got false.

Error excerpt:

ERROR: [Torch-TensorRT] - IExecutionContext::inferShapes: Error Code 7: Internal Error
... [SHUFFLE]-[aten_ops._reshape_copy.default]-[_reshape_copy]
... [GATHER]-[aten_ops.index.Tensor]-[index_index_gather]_output
Reshape dimension of -1 has no solution
Instruction: RESHAPEinput dims{16 1 1024} reshape dims{379350532 1 -1 2}

RuntimeError: [Error thrown at core/runtime/execute_engine.cpp:249]
Expected nbNames == 0 to be true but got false

Expected behavior

The compiled Torch-TensorRT engine should execute successfully and return a valid tensor with shape [B, 1, 512, 2] (for B=16) without shape-inference/internal runtime errors.

Environment

Build information about Torch-TensorRT can be found by turning on debug messages

• Torch-TensorRT Version (e.g. 1.0.0): 2.11.0
• PyTorch Version (e.g. 1.0): 2.11.0.dev20260121+cu130
• CPU Architecture: x86_64
• OS (e.g., Linux): Linux 6.14.0-29-generic (glibc 2.39)
• How you installed PyTorch (conda, pip, libtorch, source): pip (in virtualenv)
• Build command you used (if compiling from source): N/A (not built in this repro environment)
• Are you using local sources or building from archives: Torch-TensorRT is from local editable source (Editable project location: /home/wenbingl/scratch/g/t-trt)
• Python version: 3.12.3
• CUDA version: 13.0 (from torch.version.cuda)
• GPU models and configuration: 1x NVIDIA RTX PRO 6000 Blackwell Server Edition
• Any other relevant information:
  • Backend: torch.compile(..., backend="tensorrt")
  • Debug logs enabled (options["debug"] = True)
  • Repro requires dynamic shape marking (torch._dynamo.mark_dynamic)

Additional context

• The minimized repro is derived from a real decoder graph where int64 tensors are used both for indexing and symbolic shape propagation.
• In debug output, the _reshape_copy shape operand includes a symbolic value sourced from targets.size(0), and runtime appears to treat it incorrectly (garbage large integer), causing shuffle/reshape failure.
• This looks like a shape-tensor handling issue across aten.index.Tensor + _reshape_copy when integer-derived symbolic dims are involved.

[apbose]

Initial error was error in extracting symbolic shape. SymInt inputs from targets.size(0) was causing extract_symbolic_shape_expressions to return None, raising a RuntimeError during TRT compilation. Also appending
scalar inputs with an is_scalar meta data.